Risk management of chemical processes using dynamic simulation and cfd-based surrogate model approach

국문 초록 (Abstract)

공정 안전은 사고 발생시 막대한 피해로 인하여 화학 공장을 관리하는 가장 중요한 요소 중 하나로 간주되어왔다. 이러한 목표를 달성하기 위해 위험 및 운전성 연구, 보호 계층 분석, 사건 나무 분석, 나비 넥타이 분석, 그리고 정량적 위험성 평가와 같은 다양한 연구에 의해 몇가지 방법론이 도입되었다. 위에서 언급한 방법론은 일반적으로 각각의 공정 혹은 사고 정보를 제공할 수 있는 상용 소프트웨어에 의해 수행된다. 그러나 공정과 사고는 서로 밀접하게 연관되어 있으므로 두 모듈을 통합하여 정량적이고 정확하게 계산된 위험 결과를 산출하여 공정 관리의 신뢰성을 높일 필요가 있다. 따라서 본 논문에서는 공정과 사고 시뮬레이션을 통합하여 탈황 공정을 관리하는 새로운 방법론을 제안한다.
먼저, 기존의 방법으로는 감지할 수 없는 고유한 위험을 발견하기 위하여 동적 공정 시뮬레이션 및 사고 시뮬레이션과 통합된 정량적 위험성 평가 방법론이 제안된다. 이를 위해 탈황 공정에서 공정 종료 (Shut-down) 절차의 동적 공정 시뮬레이션을 실시하여 변수의 거동을 관찰하고 이를 통하여 정상상태에서 발견할 수 없었던 운전조건들을 찾아내어 정량적 위험성 평가를 수행한다.
두번째로 탈황 공정을 안전하게 관리하기 위하여 공정과 사고 시뮬레이션 결과 데이터 전송을 통해 제어실 운전원과 현장 운전원의 협력을 목표로 하는 새로운 운전원 훈련 시스템이 도입된다. 이 대화형 시뮬레이션 모듈은 화학 공장에서 발생할 수 있는 여러가지 시나리오를 미리 상정한다. 시나리오가 시작되면 제어실 운전원과 현장 운전원의 협업을 통하여 적절한 조치를 취하도록 유도되는데 이 과정에서 공정과 사고 시뮬레이션이 연동된다. 사고 결과는 전산 유체 역학을 사용하여 정확하게 계산되며 결과를 데이터베이스에 저장하여 훈련자들의 조치에 따라서 결과를 실시간으로 불러올 수 있는 구조로 되어있다.
마지막으로 막대한 용량의 데이터베이스를 대체하기 위하여 차원 축소 및 회귀 모델이 결합된 대리 모델이 제안된다. 이러한 과정을 통하여 입력 변수는 효과적으로 잠재 공간에 투영되어 효과적인 회귀 모델이 완성 되었으며 이를 통하여 전산 유체 역할을 대신할 대리 모델이 성공적으로 개발되었다.

다국어 초록 (Multilingual Abstract)

Process safety has been considered as one of the most important factors to manage a chemical plant due to a vast amount of damage in case of accidents. To achieve this goal, several methodologies have been introduced by various studies, e.g., hazard and operability (HAZOP) study, layer of protection analysis (LOPA), event tree analysis (ETA), bow-tie analysis, quantitative risk assessment (QRA), etc. The methodologies mentioned above are usually performed by commercial software which can provide each of accident or process information. However, since process and accident are closed related to each other, these two modules should be integrated to produce a quantitative and accurate risk outcome can result in the reliable process management. Thus, new methodologies to manage a heavy oil desulfurization (HOD) process by integrating process and accident simulations are proposed in this paper.
First, a methodology for quantitative risk assessment (QRA) integrated with dynamic simulation and accident simulation is proposed for the purpose of discovering inherent risks which are undetectable by the conventional method. For this, a dynamic simulation of shut-down procedure in the HOD process is performed to observe the behavior of process variables with the commercial software of Aspen HYSYS and from the dynamic simulation results, several blind spots that shows higher operating pressure than the steady-state simulation are identified. To assess the risks of the detected blind spots, QRA are performed using SAFETI. As a result of applying the proposed method to HOD process, several spots, which was identified as having low risks, turned out to be intolerably risky. In addition, mitigation procedures are carried out which reduces the risk of the process to the tolerable level, resulting in a safer and more reliable process.
Secondly, to safely operate the HOD process, a new operator training system (OTS) which targets the interactive cooperation of control room operators (CROPs) and field operators (FOPs) was introduced with the aid of data transfer between process and accident simulation results. This interactive simulation module induces the CROPs and FOPs to take proper actions in case of accident situation among pre-identified scenarios in a chemical plant. Developed model integrates the real-time process dynamic simulation by Aspen HYSYS with the off-line database of 3D-computational fluid dynamics (CFD) accident simulation results by FLACS in a designed interface using object linking and embedding (OLE) technology. As a result, an improved training effect is expected by providing accurate process and accident information to both CROPs and FOPs in real-time.
Lastly, a surrogate model which consists of dimensionality reduction and generating a regression model is introduced to replace the high-cost CFD-based FLACS model. The dimensionality reduction is performed by a variational autoencoder (VAE) combined with deep convolutional layers (DN). Through this method, the CFD results are compressed in the latent space by a probabilistic encoder. In sequence, input variables are mapped to the latent space with the aid of a deep neural network (DNN), resulting in an efficient regression model. Then this regression model is reconstructed by a probabilistic decoder, to successfully substitute the original CFD image.

목차 (Table of Contents)

• Abstract 1
• Contents 4
• List of figures 7
• List of tables 9
• Chapter 1. Introduction 10
• 1.1. Research motivation 10
• 1.2. Target process description 11
• 1.3. Outline of the thesis 13
• Chapter 2. Multi-unit dynamic simulation with 2D accident simulation: Implementation of quantitative risk assessment 14
• 2.1. Introduction 14
• 2.2 Background 16
• 2.3. Methodology 20
• 2.3.1. General sequence of QRA 20
• 2.3.2. QRA with dynamic simulation of shut-down procedure 23
• 2.4. Implementation of QRA for reactor section in HOD process 25
• 2.4.1. Target process description: Reactor section in HOD process 25
• 2.4.2. Hazard identification and dynamic simulation of shut-down procedure 30
• 2.4.3. Consequence analysis 35
• 2.4.4. Frequency analysis 38
• 2.5. Results and discussion 42
• 2.5.1. Implementation of QRA by commercial software and analysis of the results 42
• 2.5.2. Risk reduction 55
• 2.6. Conclusion 58
• Chapter 3. Plant-wide dynamic simulation with 3D accident simulation: Development of operator training system 60
• 3.1. Introduction 60
• 3.2. Background 62
• 3.3. Methodology 64
• 3.3.1. Interactive simulation technology 64
• 3.3.2. Dynamic process simulation 67
• 3.3.3. Accident simulation 70
• 3.4. Construction of OTS platform: Entire HOD process 73
• 3.4.1. Target process modeling 73
• 3.4.2. Scenario generation: From process anomaly to accident 78
• 3.4.3. Interactive simulation platform 83
• 3.5. Conclusion 85
• Chapter 4. Surrogate model construction for real-time application of accident results: Variational autoencoder with convolutional neural network 86
• 4.1. Introduction 86
• 4.2. Background 87
• 4.3. AVR-based OTS platform and jet fire CFD model 90
• 4.3.1. Architecture of the proposed OTS with AVR 90
• 4.3.2. Target accident model description 93
• 4.3.3. Mathematical formulation of jet fire 96
• 4.3.4. Numerical setup and data preparation 98
• 4.4. Surrogate model construction by VAEDC-DNN model 101
• 4.4.1. VAEDC-DNN model description 102
• 4.4.2. Numerical setting 108
• 4.5. Results and discussion 110
• 4.5.1. Performance evaluation 110
• 4.5.2. Application to AVR-based OTS 114
• 4.6. Conclusion 117
• Chapter 5. Concluding remarks 118
• Acknowledgement 119
• References 120
• Nomenclature 125
• Abstract in Korea (국문초록) 127